Gemini 4 Neon and Inkling 2 Mini Claim Near-Frontier Scores at Lower Cost, but Both Trail on Some Cited Benchmarks
Google's Gemini 4 Neon and Thinking Machines' Inkling 2 Mini both launched this month pitched as near-frontier models at lower cost, but the benchmark tables cited for each show wins, losses and a flat tie.
Background
Over the past two weeks, two AI labs have released models pitched the same way: smaller or cheaper than a sibling or rival, yet close enough in measured performance to make the price difference look like pure upside. Google launched Gemini 4 Neon on November 18, a mid-tier model that VentureBeat reports slots between the cheaper, high-volume Gemini 4 Hydrogen and the larger Gemini 4 Carbon: “Google now has an answer for three different budgets: Hydrogen for high-volume work, Neon for harder jobs that recur often, and Carbon for extended projects.” Two days later, Thinking Machines Lab released Inkling 2 Mini, a 26.4-billion-parameter model that is roughly 4.5% the size of the 594-billion-parameter Inkling 2 Small the company launched two weeks earlier.
Both launches follow a pattern the Herald has already tracked this month. Alibaba’s Qwen4-Max, covered previously, was marketed on cost reduction while an independent benchmark placed it well behind several rivals it was compared against. Neon and Inkling 2 Mini repeat the structure in a different direction: the cost or size claim is accompanied by a benchmark table, drawn partly from the vendor and partly from independent trackers, that shows the cheaper or smaller model is not uniformly ahead or uniformly behind — it wins on some measures, loses on others, and ties exactly on at least one.
Current situation
Gemini 4 Neon is priced, according to Google’s own announcement, at “$3 per million input tokens ($0.15 per million cached input tokens, a 95% discount) and $15 per million output tokens,” with a one-million-token context window. VentureBeat frames that against Anthropic’s Claude Opus 5.5, reporting that “Neon costs 25% below Anthropic’s larger, flagship Opus 5.5’s standard rates of $4 and $20 despite scoring within about a percentage point on that larger model on two engineering benchmarks,” according to VentureBeat.
The benchmark table VentureBeat reproduces does show Neon ahead of Opus 5.5 on several measures: 61 against 59 (rounded) on the Artificial Analysis Intelligence Index at Neon’s “xhigh” reasoning setting, which Google’s own announcement describes as adding “deeper, multi-step reasoning for the hardest engineering, analytical, and scientific workloads”; 56.3% against 42.5% on Zapier’s AutomationBench; and 70.5% against 63.3% on Terminal-Bench Science. The independent Vals Index, queried directly, confirms Neon’s standing: Claude Opus 5.5 scored 70.14% and placed fifth in the rankings, while Gemini 4 Neon scored 70.87% and ranked second overall, according to Vals AI’s own benchmark page — figures that track closely with the 70.9% and 70.14% VentureBeat reported.
But on two of the benchmarks in the same table, Neon trails the more expensive model: 46.2% against Opus 5.5’s 47.3% on Cognition’s FrontierCode 2.0, and 65.2% against Opus 5.5’s 65.8% on Terminal-Bench 4.1, per the same VentureBeat benchmark table. Those two gaps — 1.1 and 0.6 percentage points, both calculations the Herald’s own, based on the reported scores — are plausibly what VentureBeat means when it says Neon scores “within about a percentage point” of Opus 5.5 “on two engineering benchmarks,” though the article does not explicitly name which two benchmarks that phrase refers to.
Inkling 2 Mini presents a similar split against a same-size comparison point rather than a larger sibling. According to Thinking Machines’ own announcement, the model is “a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping,” using an architecture in which, per VentureBeat’s description, “the first six layers (the prelude) and the final six layers run once, while the middle 12 layers run three times using shared weights.” Thinking Machines’ own page corroborates the structure: “Blocks 2 and 3 share weights across three passes, giving 48 layer applications per token from 24 stored layers.”
On the Artificial Analysis Intelligence Index, Inkling 2 Mini and the comparably sized Qwen3.8-27B land on an exact tie: both score 20.1, per VentureBeat’s reporting of the model’s own announcement, a result Artificial Analysis’s own writeup confirms independently, stating “Thinking Machines has released Inkling 2 Mini at 20 on the Artificial Analysis Intelligence Index.” Around that tie, Thinking Machines and Artificial Analysis both report a large efficiency gap in Inkling 2 Mini’s favor: Artificial Analysis’s own page states “Inkling 2 Mini is ~6x cheaper and ~4x faster per task than Qwen3.8 27B,” with the specific figures that “it costs $0.59 and takes 9.0 minutes per Intelligence Index task vs. $3.53 and 38.2 minutes,” and that on Artificial Analysis’s compute proxy, Inkling 2 Mini “sits on the Pareto frontier, using ~2.8x less compute than Qwen3.8 27B.” VentureBeat’s independent report of the same comparison matches those figures exactly.
The tie does not hold everywhere. On a factual-recall benchmark, Thinking Machines’ own disclosure — as reported by VentureBeat — shows Inkling 2 Mini trailing: “Inkling 2 Mini scores -8.6 on AA-Omniscience v1.1 (with a 63.4% non-hallucination rate), trailing denser peers like Qwen3.8-27B (-6.4 and 71.4%).” Inkling 2 Mini ships under an Apache 2.0 license with a standard Model Acceptable Use Policy, and is priced, per VentureBeat, at $0.15 per million input tokens and $0.50 per million output tokens through Thinking Machines’ first-party API, or $0.08 and $0.35 on OpenRouter — considerably cheaper than the $0.30/$0.03/$1.00 Tinker pricing the Herald previously reported for the larger Inkling 2 Small.
Implications
Neither company’s headline framing is false on its own terms. Neon genuinely costs less than Opus 5.5 and genuinely beats it on several of the cited benchmarks; Inkling 2 Mini genuinely matches Qwen3.8-27B’s headline index score at a fraction of the compute, speed and cost. What the fuller tables show, once assembled from the vendor pages, the independent trackers and the reporting layered on top, is that “near-frontier performance at lower cost” compresses several different comparisons — some favorable to the cheaper model, some not — into a single marketing framing built around whichever benchmarks make the gap look smallest or the lead look widest.
That compression is now a recurring feature of model launches rather than a one-off. The Herald’s earlier analysis of Qwen4-Max found the same structure in reverse: a vendor-selected comparison chart showing broad improvement, next to an independent index showing the model trailing several rivals it was benchmarked against. Readers comparing Neon, Inkling 2 Mini or any future release on a single advertised number — a percentage-point gap, a multiplier on cost or speed — are, in each of these cases, seeing one row of a longer table in which the same two models trade places depending on which benchmark is asked.
What remains unresolved in both cases is how those benchmark tables translate into real-world task performance beyond the specific evaluations cited. Vals Index, Artificial Analysis, FrontierCode, Terminal-Bench and AA-Omniscience each measure different things — professional-task accuracy, general intelligence proxies, coding tasks, agentic terminal use and factual recall, respectively — and none of the sources reviewed here claims its benchmark is a substitute for the others. The cited sources do not report how Neon or Inkling 2 Mini perform on tasks outside the specific benchmarks named above.